BUSINESSNew tool · · 2 min read · Source: TechCrunch
Box and Nvidia: AI retrieval aims to cut agent tokens by 90%
At its BoxWorks conference on November 5, 2026, Box unveiled a new retrieval layer for AI agents built on Nvidia's Nemotron models. Instead of whole documents, agents get only the relevant passages. It arrives in early 2027 for customers on the Enterprise Advanced plan.
Related tool · AdRunable – AI agent for finished workTry it free
Key facts
- On November 5, 2026, at BoxWorks, Box unveiled AI retrieval for agents built on Nvidia's Nemotron models.
- Box and Nvidia claim about 20% better accuracy, about 90% fewer tokens and roughly 4x faster answers; not independently verified.
- Launch in early 2027 for Box Enterprise Advanced customers; Nemotron 3 Ultra is already available to them in Box AI.
What Box announced
Box's press release calls the feature Box Context Retrieval (TechCrunch refers to it as "Content Retrieval"). It uses Nvidia's open Nemotron embedding models to capture what files mean, not just which keywords they contain, and Box says this extends to scanned documents, images, audio and video.
When an AI agent asks a question, it no longer has to read every document in full. It receives the passages that match the query. Existing permissions apply, so an agent only sees what the person it works for is allowed to see.
The numbers (vendor claims)
- about 20% more accurate results
- about 90% fewer tokens used
- answers roughly four times faster
These figures come from Box and Nvidia and have not been independently verified. Box uses law firms as an example: finding relevant clauses and precedents faster, even when they are worded differently.
Availability
According to Box, the retrieval feature arrives in early 2027 for Enterprise Advanced customers. Nvidia's Nemotron 3 Ultra model is already available to them in Box AI. Box did not give separate pricing. The two companies are also connecting Nvidia's OpenShell secure runtime to Box so agent workflows can be supervised and bound to policies.
What this means for you
- AI agents get cheaper when they receive the right passages instead of entire folders.
- Permissions become the foundation: agents only see what the person is allowed to see.
- For now, this is for larger companies on Box Enterprise Advanced, not individual users.
How to use this
If your company uses AI agents with files, start by reviewing permissions in your storage. New agent retrieval inherits them, so wrong sharing settings become an AI problem too.
Luigi's take
The announcement sounds technical, but it touches a question many companies are asking right now: what does AI cost us when agents keep reading entire documents? If good retrieval really cuts usage that much, it makes agents economical for many more teams. The 90% is a vendor number, though, and I would only trust it once it holds up in your own tests.
My advice for companies: whether you use Box, Dropbox or SharePoint, get your permissions and folder structure right. That is what these new agent features build on. Teams that clean this up now can switch such tools on quickly and with confidence later.
Source
TechCrunch: Box taps Nvidia's Nemotron for new agent-first content retrieval system
More sources
- Investing.com – Zusammenfassung der Box-Pressemitteilung zur BoxWorks 2026
This article was researched and written with AI assistance. Responsible: Luigi Lugmayr.
